🎞️ AI-Generated Video Detector

Mixed-domain supervised-contrastive R(2+1)D, evaluated with a fixed real/fake prototype protocol in projection space. Model & weights: jai1th/dvf_trained_transferred_aegis.

⚠️ Research demo. Built from fixed support-bank prototypes — it is not a calibrated or deployment-grade detector. It targets end-to-end AI-generated video only (e.g. Sora / KLing / Pika–style clips), not face-swap deepfakes. Do not use for moderation, legal, or forensic decisions.

Confidence is a temperature-scaled softmax over the real/fake prototype similarities (research demo — indicative, not a calibrated probability). Preprocessing matches training exactly: decode → resize short side to 224 → center-crop square → 24 frames @ 24 fps, pixels in [0, 1], no Kinetics normalisation.